Skip to content

perf-loop.md

Latest commit

 

History

History
51 lines (34 loc) · 2.17 KB

perf-loop.md

File metadata and controls

51 lines (34 loc) · 2.17 KB

Performance Loop Runner

Use the PowerShell runner at scripts/run_perf_loop.ps1 to capture a full local optimization cycle with stable artifacts under target/perf-loops/.

Commands

Initial pass:

powershell -ExecutionPolicy Bypass -File .\scripts\run_perf_loop.ps1 -CycleLabel rpc-pass-01

After making the optimization change, resume from the verification half of the same cycle:

powershell -ExecutionPolicy Bypass -File .\scripts\run_perf_loop.ps1 -CycleLabel rpc-pass-01 -ResumeFromOptimize

Dry-run orchestration without invoking cargo or python:

powershell -ExecutionPolicy Bypass -File .\scripts\run_perf_loop.ps1 -CycleLabel smoke-check -DryRun

Optional benchmark filter:

powershell -ExecutionPolicy Bypass -File .\scripts\run_perf_loop.ps1 -CycleLabel queue-pass-01 -BenchFilter tier3_system_queue

Output

Each cycle writes to target/perf-loops/<cycle-label>/:

  • manifest.json: phase status, commands, snapshot paths, selected optimization target, comparison result
  • logs/: raw command output for each test, bench, and summary phase
  • snapshots/baseline/: first benchmark summary, perf scorecard, and copied benchmark artifacts
  • snapshots/verification/: second benchmark summary, perf scorecard, and copied benchmark artifacts after optimization
  • comparison.json: before/after improvement summary once verification is complete

Selection Rules

The runner records an optimization checkpoint after the first test and benchmark pass. It selects one candidate deterministically using:

  • config/perf_targets.json as the source of truth
  • target-bucket order: engine_core, then service_budget/direct_api, then service_budget/transport, then service_budget/contention, then internal_explainer
  • percent over the operational_target inside each bucket
  • percent over the stretch_target as the next tie-breaker inside the bucket
  • higher current mean_us as the final tie-breaker inside the bucket

The runner does not mutate source code during optimization. It records the target, then expects the engineer to make the code change before running the resume step.